Generative AI models learn a distribution over content (tokens, pixels, audio frames) and sample new outputs—unlike classifiers that only assign labels.
Generative vs discriminative
| Discriminative | Generative |
|---|---|
| Predicts label from input (spam/not spam) | Produces new content conditioned on input |
| Logistic regression, classic BERT classifier | GPT-style LLMs, diffusion image models |
Modalities you will meet
- Text — chat, summarization, code completion
- Images — text-to-image, inpainting
- Audio / speech — TTS, transcription
- Multimodal — image + text in one conversation
Foundation models
A foundation model is pretrained on broad data then adapted (prompting, RAG, fine-tuning) to many tasks. Capability scales with data, compute, and architecture—but also risk surface.
Important interview questions and answers
- Q: Is autocomplete generative?
A: Yes—next-token prediction is generative; the model samples from a probability distribution over vocabulary. - Q: Does generative imply correct?
A: No—fluency ≠ factuality; evaluation and grounding are separate engineering work.
Self-check
- Give one discriminative and one generative product example.
- Why is fluency dangerous without citations?
Tip: When stakeholders say "AI," ask if they mean classification, ranking, or generation—scope changes architecture.
Interview prep
- Generative vs discriminative?
Discriminative labels inputs; generative samples new content from learned distributions.
- Foundation model?
Large pretrained model adapted via prompts, RAG, or fine-tuning to many downstream tasks.